Germany has several local AI safety groups, but AI safety lacks a national presence. SAIGE could be a key organization in tying these groups together to create a more coordinated German approach to making AI safer.

Germany has raw resources that could be used to make an impact in AI safety, but needs more field building efforts like these to tap into that potential.

Germany needs a stronger AI safety community and this is the strongest effort I'm currently aware of!

@Corn Thanks Cornelis! And conversely I'm grateful at how you bring ideas like Taco Tuesday back and share them with your Canada friends. Bring them to Mox sometime!

(also: what are the other 2??)

@ollive Thanks Carolina! Really appreciate your recent volunteering with Mox, decorating offices and events, and just the spirit of exuberance you embody every single time I run into you

What progress have you made since your last update?

We have released TransformerLens 3 beta! Second beta release is coming up tomorrow. We decided to keep the project in beta until all features of TransformerLens 2 are ready. Right now 90% of use cases are covered, and we have hypothetical support for almost 5k models, compared to TransformerLens 2 200.

What are your next steps?

We are continuing iterating on TransformerLens 3 until we cover absolutely all use cases from the old version of TransformerLens, at which point we will have an official TransformerLens 3.0.

Is there anything others could help you with?

Full time development has been funded through the end of the year!

AI safety is a rapidly growing global concern, especially with the rapid commercial development of models with seemingly little concern for safeguarding; however, I believe that with good guardrails, AI holds immense potential for innovation. This proposal identifies a clear gap: a mismatch between potential talent in Germany and a lack of infrastructure to encourage this talent towards this emerging and important field of AI safety. I have faith in Jess, I sincerely hope that SAIGE succeeds because it would be a shame otherwise, as the statement of risk makes clear.

Full disclosure: I know Jess personally and she's always shown a relentless drive and immense conscientiousness.

This is filling a gap in the German market and needs to succeed!

Thank you @Ashe ! SAIGE is exactly trying to fix the issue you have described. The incubator is a first step (and so excited to look through your project proposal!). We also *just* confirmed with Successif that the first career workshop is happening, on March 11th. It'll be added to the SAIGE calendar very soon.

The MoSSAIC framework is good foundational work that pushes mechanistic interpretability in a more scientifically grounded and principled direction. I recommend funding this work. This paper provides a meta-level intervention while identifies a core assumption underlying most contemporary safety work (the "causal-mechanistic paradigm") and systematically demonstrates why this paradigm will likely fail as AI systems become more capable. It does this not only through abstract speculation alone, but by connecting concrete empirical results (Bailey et al. 2024 on obfuscated activations, McGrath et al. 2023 on self-repair) to a sequence of plausible threat models. I have high respect for Matt Farr for taking the initiative to work on something intrinsically valuable to the Alignment field yet that is misaligned with the incentives of other main players. I believe that this is a diagnostic and constructive work that questions the dominant paradigm that the field is desperately needed yet not being talked about nor discussed enough. The field systematically underfunds this kind of work because it does not produce legible outputs but the MoSSAIC framework is an example of a legible output. I encourage people with more resources to retroactively fund this work.

This is funding for a project Aria did with me during MATS, but before the official program start/before she had access to her MATS compute budget. I think this was an exciting project, benchmarking a bunch of ideas for how to shift how RL generalises, since I haven't seen this done well before, and ensuring RL generalises correctly is a core alignment sub problem. The amount is high because RL is compute intensive. This is a mild conflict of interest, but my bar from mentoring someone is significantly higher than my bar for funding them, so I do not expect that this is warping my judgements

What progress have you made since your last update?

Since my last update, I completed the seed-funded scoping work to assess Liberia’s biosafety and biosecurity landscape and to define an evidence-based pathway for strengthening this capacity in alignment with IHR (2005) and the WHO-led JEE 2023 findings.

I conducted a BLAM-based scoping assessment using a mixed-method approach, including a targeted desk review, key informant consultations, and structured facility assessments. I applied and refined BLAM for the Liberia context, then synthesized the evidence into domain scoring and risk-based prioritization to identify the most consequential gaps and immediate actions.

I consolidated and documented key findings highlighting priority weaknesses in governance and oversight, operational biosafety practices, biosecurity accountability mechanisms (including access control and inventory systems), equipment maintenance and certification systems, waste management and decontamination, and safe specimen handling and referral pathways. I also held stakeholder validation and coordination discussions with national counterparts and technical stakeholders to ensure the findings reflect operational realities and to support a shared understanding of priorities.

Finally, I developed a clear pathway for scale-up, including a scoping brief outlining practical options for strengthening the biosafety and biosecurity legal/regulatory and oversight environment, and an investable outline for a follow-on implementation phase (12–24 months) with priority workstreams, indicative sequencing, and resource needs. 

What are your next steps?

My next step is to translate the scoping findings into an expanded implementation phase that addresses the priority gaps identified through BLAM. Building on the evidence generated during this seed project, I will work with NPHIL and the Ministry of Health to agree on a phased set of recommendations that can be taken forward as a larger, more detailed programme of biosafety and biosecurity strengthening for epidemic preparedness and IHR implementation.

A key outcome of this seed phase is that it has attracted strong attention from NPHIL and generated institutional buy-in to move beyond assessment toward action. I will therefore support NPHIL’s leadership to use the BLAM findings as a practical decision tool, including aligning partners, defining roles, and sequencing priority interventions in a way that is feasible within Liberia’s system constraints.

In practical terms, I plan to expand the scope by: (i) broadening BLAM application to additional laboratory tiers and relevant sectors, (ii) advancing the agreed governance and oversight pathway, including options for strengthening legislation and regulatory mechanisms, and (iii) developing a more detailed, costed implementation plan that can be used to secure follow-on funding and guide coordinated execution of the highest-priority actions.

Is there anything others could help you with?

Yes. To move from this completed scoping phase to a full implementation program with measurable improvements in Liberia’s biosafety and biosecurity capacity, targeted support from funders and technical collaborators would be highly valuable in three areas:

1) Follow-on financing for an implementation phase (12–24 months).
Support is needed to operationalize the BLAM recommendations through a phased program that can strengthen governance and oversight, establish routine certification and maintenance systems, improve waste and decontamination practices, and expand workforce capacity. Flexible funding that can combine coordination, technical assistance, and selected high-impact investments would enable rapid transition from planning to action.

2) Technical collaboration and specialized expertise.
Partners could contribute expertise in laboratory biosafety and biosecurity systems (including risk assessment, incident management, and biosecurity accountability), equipment certification/maintenance systems (e.g., biosafety cabinets and decontamination validation), safe waste management solutions, and monitoring frameworks aligned with IHR and JEE. Collaboration to refine BLAM as a repeatable national monitoring tool and to embed it in routine supervision would also strengthen sustainability.

3) Implementation partnerships and capacity building platforms.
Support from institutions that can co-deliver training, mentorship, and supportive supervision—particularly through NPHIL/MoH structures and platforms such as the national laboratory network and FETP would help scale competency and institutionalize practices. In addition, partners that can help convene and coordinate through NEPC will accelerate alignment and reduce duplication of efforts.

Overall, additional support would help Liberia convert strong national buy-in and a completed evidence base into a coordinated, scalable program with direct relevance to epidemic preparedness and global health security.

All the best from Canada! It has been inspiring seeing the community-building you are working towards. Hope it becomes a template for more spaces like it to pop up in other areas around the world!

This project has potential for enabling copius amounts of AI safety research that counterfactually wouldn't happen. The current lack of infrastructure around the field in Germany causes some to move away to pursue work abroad. However, there are many others who simply "move on" to other work they can do where they already live, or are never informed that AI safety or alignment is an option. SAIGE could help address those issues.

In the short term, I'm optimistic about the incubator program both introducing new people to AI alignment work and eventually enabling them to continue it. I have applied to the program as a mentor for projects in agent foundations related to my research at MATS.

@Austin, thanks! As for the web UI, I just started working on HTML logs yesterday. I also quite like the idea of a human-in-the-loop.

I'm also in the midst of a refactor that makes for much easier extensions to different types of games. I've been watching The Traitors as of late, so Mafia is a game I'm especially excited to implement.

Mox is currently one of my top 3 fave places on the entire earth, unironically. Thank you for everything you do and everything you stand for! 🧡

Awesome space.

Thanks a bunch @Yannick-Muehlhaeuser 😇

Thank you @mk_lao ! It was great seeing you at Mox, and let's keep in touch ;)

Great to finally see such a national organization forming. Germany really needs a platform helping STEM researchers organize, learn and build.

I've been incredibly impressed by Jessica's agentiness and drive. The team she has on board is great and she is quite well coordinated with the relevant other stakeholders in Germany and Europe. Given the clear plan for the future, I would be quite excited if we could get this off the ground.

Germany has several local AI safety groups, but AI safety lacks a national presence. SAIGE could be a key organization in tying these groups together to create a more coordinated German approach to making AI safer.

Germany has raw resources that could be used to make an impact in AI safety, but needs more field building efforts like these to tap into that potential.

Germany needs a stronger AI safety community and this is the strongest effort I'm currently aware of!

@Corn Thanks Cornelis! And conversely I'm grateful at how you bring ideas like Taco Tuesday back and share them with your Canada friends. Bring them to Mox sometime!

(also: what are the other 2??)

@ollive Thanks Carolina! Really appreciate your recent volunteering with Mox, decorating offices and events, and just the spirit of exuberance you embody every single time I run into you

What progress have you made since your last update?

We have released TransformerLens 3 beta! Second beta release is coming up tomorrow. We decided to keep the project in beta until all features of TransformerLens 2 are ready. Right now 90% of use cases are covered, and we have hypothetical support for almost 5k models, compared to TransformerLens 2 200.

What are your next steps?

We are continuing iterating on TransformerLens 3 until we cover absolutely all use cases from the old version of TransformerLens, at which point we will have an official TransformerLens 3.0.

Is there anything others could help you with?

Full time development has been funded through the end of the year!

AI safety is a rapidly growing global concern, especially with the rapid commercial development of models with seemingly little concern for safeguarding; however, I believe that with good guardrails, AI holds immense potential for innovation. This proposal identifies a clear gap: a mismatch between potential talent in Germany and a lack of infrastructure to encourage this talent towards this emerging and important field of AI safety. I have faith in Jess, I sincerely hope that SAIGE succeeds because it would be a shame otherwise, as the statement of risk makes clear.

Full disclosure: I know Jess personally and she's always shown a relentless drive and immense conscientiousness.

This is filling a gap in the German market and needs to succeed!

Thank you @Ashe ! SAIGE is exactly trying to fix the issue you have described. The incubator is a first step (and so excited to look through your project proposal!). We also *just* confirmed with Successif that the first career workshop is happening, on March 11th. It'll be added to the SAIGE calendar very soon.

The MoSSAIC framework is good foundational work that pushes mechanistic interpretability in a more scientifically grounded and principled direction. I recommend funding this work. This paper provides a meta-level intervention while identifies a core assumption underlying most contemporary safety work (the "causal-mechanistic paradigm") and systematically demonstrates why this paradigm will likely fail as AI systems become more capable. It does this not only through abstract speculation alone, but by connecting concrete empirical results (Bailey et al. 2024 on obfuscated activations, McGrath et al. 2023 on self-repair) to a sequence of plausible threat models. I have high respect for Matt Farr for taking the initiative to work on something intrinsically valuable to the Alignment field yet that is misaligned with the incentives of other main players. I believe that this is a diagnostic and constructive work that questions the dominant paradigm that the field is desperately needed yet not being talked about nor discussed enough. The field systematically underfunds this kind of work because it does not produce legible outputs but the MoSSAIC framework is an example of a legible output. I encourage people with more resources to retroactively fund this work.

This is funding for a project Aria did with me during MATS, but before the official program start/before she had access to her MATS compute budget. I think this was an exciting project, benchmarking a bunch of ideas for how to shift how RL generalises, since I haven't seen this done well before, and ensuring RL generalises correctly is a core alignment sub problem. The amount is high because RL is compute intensive. This is a mild conflict of interest, but my bar from mentoring someone is significantly higher than my bar for funding them, so I do not expect that this is warping my judgements

What progress have you made since your last update?

Since my last update, I completed the seed-funded scoping work to assess Liberia’s biosafety and biosecurity landscape and to define an evidence-based pathway for strengthening this capacity in alignment with IHR (2005) and the WHO-led JEE 2023 findings.

I conducted a BLAM-based scoping assessment using a mixed-method approach, including a targeted desk review, key informant consultations, and structured facility assessments. I applied and refined BLAM for the Liberia context, then synthesized the evidence into domain scoring and risk-based prioritization to identify the most consequential gaps and immediate actions.

I consolidated and documented key findings highlighting priority weaknesses in governance and oversight, operational biosafety practices, biosecurity accountability mechanisms (including access control and inventory systems), equipment maintenance and certification systems, waste management and decontamination, and safe specimen handling and referral pathways. I also held stakeholder validation and coordination discussions with national counterparts and technical stakeholders to ensure the findings reflect operational realities and to support a shared understanding of priorities.

Finally, I developed a clear pathway for scale-up, including a scoping brief outlining practical options for strengthening the biosafety and biosecurity legal/regulatory and oversight environment, and an investable outline for a follow-on implementation phase (12–24 months) with priority workstreams, indicative sequencing, and resource needs. 

What are your next steps?

My next step is to translate the scoping findings into an expanded implementation phase that addresses the priority gaps identified through BLAM. Building on the evidence generated during this seed project, I will work with NPHIL and the Ministry of Health to agree on a phased set of recommendations that can be taken forward as a larger, more detailed programme of biosafety and biosecurity strengthening for epidemic preparedness and IHR implementation.

A key outcome of this seed phase is that it has attracted strong attention from NPHIL and generated institutional buy-in to move beyond assessment toward action. I will therefore support NPHIL’s leadership to use the BLAM findings as a practical decision tool, including aligning partners, defining roles, and sequencing priority interventions in a way that is feasible within Liberia’s system constraints.

In practical terms, I plan to expand the scope by: (i) broadening BLAM application to additional laboratory tiers and relevant sectors, (ii) advancing the agreed governance and oversight pathway, including options for strengthening legislation and regulatory mechanisms, and (iii) developing a more detailed, costed implementation plan that can be used to secure follow-on funding and guide coordinated execution of the highest-priority actions.

Is there anything others could help you with?

Yes. To move from this completed scoping phase to a full implementation program with measurable improvements in Liberia’s biosafety and biosecurity capacity, targeted support from funders and technical collaborators would be highly valuable in three areas:

1) Follow-on financing for an implementation phase (12–24 months).
Support is needed to operationalize the BLAM recommendations through a phased program that can strengthen governance and oversight, establish routine certification and maintenance systems, improve waste and decontamination practices, and expand workforce capacity. Flexible funding that can combine coordination, technical assistance, and selected high-impact investments would enable rapid transition from planning to action.

2) Technical collaboration and specialized expertise.
Partners could contribute expertise in laboratory biosafety and biosecurity systems (including risk assessment, incident management, and biosecurity accountability), equipment certification/maintenance systems (e.g., biosafety cabinets and decontamination validation), safe waste management solutions, and monitoring frameworks aligned with IHR and JEE. Collaboration to refine BLAM as a repeatable national monitoring tool and to embed it in routine supervision would also strengthen sustainability.

3) Implementation partnerships and capacity building platforms.
Support from institutions that can co-deliver training, mentorship, and supportive supervision—particularly through NPHIL/MoH structures and platforms such as the national laboratory network and FETP would help scale competency and institutionalize practices. In addition, partners that can help convene and coordinate through NEPC will accelerate alignment and reduce duplication of efforts.

Overall, additional support would help Liberia convert strong national buy-in and a completed evidence base into a coordinated, scalable program with direct relevance to epidemic preparedness and global health security.

All the best from Canada! It has been inspiring seeing the community-building you are working towards. Hope it becomes a template for more spaces like it to pop up in other areas around the world!

This project has potential for enabling copius amounts of AI safety research that counterfactually wouldn't happen. The current lack of infrastructure around the field in Germany causes some to move away to pursue work abroad. However, there are many others who simply "move on" to other work they can do where they already live, or are never informed that AI safety or alignment is an option. SAIGE could help address those issues.

In the short term, I'm optimistic about the incubator program both introducing new people to AI alignment work and eventually enabling them to continue it. I have applied to the program as a mentor for projects in agent foundations related to my research at MATS.

@Austin, thanks! As for the web UI, I just started working on HTML logs yesterday. I also quite like the idea of a human-in-the-loop.

I'm also in the midst of a refactor that makes for much easier extensions to different types of games. I've been watching The Traitors as of late, so Mafia is a game I'm especially excited to implement.

Mox is currently one of my top 3 fave places on the entire earth, unironically. Thank you for everything you do and everything you stand for! 🧡

Awesome space.

Thanks a bunch @Yannick-Muehlhaeuser 😇

Thank you @mk_lao ! It was great seeing you at Mox, and let's keep in touch ;)

Great to finally see such a national organization forming. Germany really needs a platform helping STEM researchers organize, learn and build.

I've been incredibly impressed by Jessica's agentiness and drive. The team she has on board is great and she is quite well coordinated with the relevant other stakeholders in Germany and Europe. Given the clear plan for the future, I would be quite excited if we could get this off the ground.